Applications of Artificial Neural Networks in Microorganism Image Analysis: A Comprehensive Review from Conventional Multilayer Perceptron to Popular Convolutional Neural Network and Potential Visual Transformer

1 Introduction

• Manual morphological analysis is cost-effective but laborious, subjective, and time-consuming, while noise, filopodia, lighting, transparency, and varied tasks challenge robust automation.AI can automate laborious image work and reduce subjective differences in microorganism analysis.Source: §1 · pp. 2–3

1.1 Artificial Intelligence Methods for Microorganism Image Analysis

• MIA uses conventional machine learning and classical or deep ANNs; unlike feature-engineered classical workflows, CNNs learn image features directly and have gained advantages in MIA.Classical ANN examples include MLP, RBF, and PNN; deep-network examples include U-Net segmentation and Xception-based bacterial classification.Source: §1.1 · p. 5

1.2 Motivation of This Review

• Unlike prior reviews centered on clustering, segmentation, classification, or counting, this review centers ANN applications across seven MIA tasks.It analyzes 96 papers from 1992 to 2020 and provides representative-network introductions, method discussion, summary tables, statistical analysis, and potential directions.Source: §1.2 · pp. 6–7

1.3 Paper Searching and Screening

• 95 research papers were retained; the review describes a shift from feature-engineered classical ANNs before 2014 to deep ANNs, especially CNNs, afterward.The search combined prior-review papers with database searches using MIA and neural-network keywords, followed by screening and additions.Source: §1.3 · pp. 7–8, Figure 4 · p. 8

1.4 Structure of This Review

• The paper proceeds from representative ANN development to classical and deep ANN applications, then method analysis, potential directions, and conclusions.These topics are organized across Sections 2–6.Source: §1.4 · p. 8

2.1 Evolution of Artificial Neural Networks

• ANN history is framed as three stages: M-P neuron and perceptron sparked an early rise, but perceptron’s inability to solve XOR led to a research trough.The BP algorithm [134] enabled MLPs to solve XOR and train CNNs, though early training was slow and prone to overfitting. Later improvements in computing and data availability helped ANN research rise again, with advantages in speech and image recognition.Source: §2.1 · pp. 9–10

2.2 Representative Artificial Neural Networks

• Classical MIA commonly used feature engineering plus classifiers, especially MLP, while CNNs became popular for classification (AlexNet, VGGNet [145], ResNet, Inception), segmentation (U-Net), and detection (YOLO).An MLP has input, hidden, and output layers; back-propagation lets it adjust connection weights layer by layer.Source: §2.2 · p. 10, §2.2.1 · p. 10

2.2.2 AlexNet

• AlexNet won ILSVRC 2012 with eight layers, using ReLU, multi-GPU training, overlapping pooling, data augmentation, and dropout.It has 60 million parameters.Source: §2.2.2 · p. 11

2.2.3 VGGNet

• VGGNet [145] uses stacked 3 × 3 filters to capture spatial information with fewer parameters than larger filters, enabling deeper networks.VGG-16 and VGG-19 have 138 and 144 million parameters, respectively [145].Source: §2.2.3 · p. 11

2.2.4 Inception

• GoogLeNet [149] introduced Inception, combining multiple filter sizes for multi-level features and 1 × 1 convolutions to reduce parameters.It won ILSVRC 2014 with 22 convolutional layers and 7 million parameters; later versions added batch normalization, factorized filters, or residual blocks.Source: §2.2.4 · p. 13

2.2.5 ResNet

• ResNet uses identity mapping and residual blocks to deepen networks while addressing gradient vanishing or explosion associated with simply adding layers.The residual function is F(x) = H(x) − x.Source: §2.2.5 · p. 13

2.2.6 U-Net

• U-Net was initially used for microscopic image segmentation and pairs contracting and expansive paths in a symmetric architecture.Its training uses strong data augmentation to make better use of annotated samples, and its end-to-end structure retrieves shallow information.Source: §2.2.6 · p. 13

2.2.7 YOLO

• YOLO divides an image into an S × S grid, with each cell predicting bounding boxes, confidence scores, and conditional class probabilities.Its architecture has 24 convolutional layers and two fully connected layers, using 1 × 1 reduction layers followed by 3 × 3 convolutions.Source: §2.2.7 · p. 14

3 Classical Neural Network for Microorganism Image Analysis

• Classification studies make up most classical-neural-network microorganism image analysis works, so the review analyzes them by the networks used.Other tasks are reviewed independently because each contains relatively few papers; the section also summarizes characteristics and provides a table to help readers find related studies.Source: §3 · p. 15

3.1.1 Classification based on MLP

• Around 30 MLP-based classification papers vary in preprocessing, features, and architectures; most segment ROIs, favor shape features, and rarely screen features, limiting efficiency.Some works select effective features using a classification tree.Source: §3.1.1 · pp. 15, 17, 19
• Three-layer BP-based MLPs predominate, with no reported model exceeding four layers; neuron counts generally depend on feature-vector dimensions and class count.Most works provide little theoretical or experimental analysis of hidden-layer settings, and some omit implementation details.Source: §3.1.1 · pp. 19–21

3.1.2 Classification based on RBF

• Most RBF methods extract features directly, but uses a t-test and a classification tree to screen features for more efficient classification.[120] selects compactness, eccentricity, circularity, and tortuosity; selects from shape, texture, and color features.Source: §3.1.2 · p. 22
• RBF outperforms MLP in and usually beats alternatives, while favors MLP; report 100% accuracy for RBF on spiral-cell classification.In, RBF, 3σ, and fuzzy classifiers each reach 100% classification accuracy; in, RBF and the neuro-fuzzy classifier do so.Source: §3.1.2 · pp. 22–23

3.1.3 Classification based on other classical ANNs

• The dual-classification system in assigns a class only when LVQ-NN and SVM agree, reducing error by 50% to 100% versus previous methods.The experiments use seven plankton categories and about 20,000 images, and the corrected system supports abundance estimation.Source: §3.1.3 · p. 23
• For TB tissue detection, reports 98.07% accuracy, 100% sensitivity, and 96.19% specificity; reaches 77.33% highest testing accuracy.For, average testing accuracy is 74.62% for 35 hidden nodes.Source: §3.1.3 · pp. 23–24
• A PNN using nine geometrical, optical, and textural features classified five microorganism categories with 100% accuracy.Features were extracted from pre-processed images.Source: §3.1.3 · p. 24
• The microalgae method combined image segmentation, shape-feature extraction, pigment-signature determination, and SOM classification.The SOM neuron count was fixed to the estimated number of categories, with neurons interconnected through neighborhood relations.Source: §3.1.3 · p. 24
• The SOM-based method achieved 98.6% accuracy across 23 categories and 53,869 images.The study used a private dataset and targeted real-time microalgae identification and enumeration.Source: §3.1.3 · p. 24

Segmentation Task

• A Self-Tunning Vision System used a SOM to segment Aureobasidium pullulans images while monitoring morphological changes during pullulan fermentation.The workflow also included preprocessing, intermediate processing, feature extraction, and classification; the review describes the system as reliable, accurate, and versatile.Source: §3.2 · p. 25
• 98.72%, 99.45%, and 97.75% accuracy were achieved by HMLP segmentation on tuberculosis tissue-slide samples A, B, and C, respectively.The HMLP uses hue and saturation from each pixel’s 3 × 3 neighborhood as inputs and classifies the center pixel; training used 5,608 data points.Source: §3.2 · p. 25
• 96.80% sensitivity with a 3.38% error rate was SVM’s best result, while ANN achieved best sensitivity of 91.53% and an error rate of 5.20%.Segmentation classifiers used pixel-color features from four color spaces with scalar feature selection, trained on 1,200 bacilli pixels and 1,200 background pixels from 120 images.Source: §3.2 · p. 25

3.3 Counting Task

• The feed-forward MLP used 81 inputs, ten hidden neurons, and one output, processing grey values from 9 × 9 pixel sub-images without feature engineering.The sub-images were taken from 512 × 512 pixel microscope images.Source: §3.3 · p. 25
• 93%-98% accuracy was reported for enumerating bacteria on metal or silicone rubber substrata in low-quality images, and 98% for yeast cells on silicone rubber.For high-contrast images, nearly all adhering organisms were detected correctly.Source: §3.3 · p. 26
• The network produced an output for every original-image pixel, after which adhering microorganisms were indicated with marking circles.The approach was proposed for image enhancement and enumeration of microorganisms adhering to solid substrates.Source: §3.3 · pp. 25–26

3.4 Feature Extraction Task

• The method addresses complex and ineffective manual bacteria classification with an image-classification approach using PCNN-generated features.PCNN is described as a biologically inspired artificial neural network based on.Source: §3.4 · p. 26
• Images are cropped to save the region of interest before PCNN generates entropy sequence features for classification.Source: §3.4 · p. 26
• The method classifies bacteria using a classifier based on Euclidean distance, and experiments found it feasible and efficient.Source: §3.4 · p. 26

• Privately collected datasets and the lack of public data make general MIA algorithms and comparative evaluation difficult.Source: §3.5 · p. 26
• Classical-neural-network MIA usually uses preprocessing, feature extraction, and classification, making feature engineering important to performance.Source: §3.5 · p. 26
• Classical neural networks in MIA are usually shallow because deep networks’ many interlayer connections can create abundant parameters and difficult training.Source: §3.5 · p. 26
• Most classical-neural-network MIA tasks are classification, although some networks directly process image data, unlike the later CNN approach.Source: §3.5 · p. 26

4 Deep Neural Network for Microorganism Image Analysis

• Deep-network microorganism image analysis spans six tasks, with classification receiving the most detailed treatment and papers organized by network type.The section also summarizes the methods and provides a table to help readers find relevant papers.Source: §4 · p. 29

4.1.1 Classification based on CNN

• A parallel CNN with motor modules recognized approximately 96% of coccoliths from 11 Pleistocene taxa during routine work.The modules learn image transformations; SYRACO2, a shallower CNN, reported an 86% average effective recognition rate across 13 species.Source: §4.1.1 · pp. 29–30
• Fusion-Net reached 98.4% classification accuracy by combining morphological, intensity, and spectral features analyzed by three networks.Before fusion, LSTM, ResNet, and 1D CNN achieved 92.2%, 93.8%, and 96.2%, respectively.Source: §4.1.1 · p. 32
• For TB classification, a feature-engineered CNN framework achieved 99.5% accuracy, 94.7% sensitivity, and 99% specificity.Other CNN bacterial classifiers reported 94.9% accuracy and 95.6% F1-score with autoencoder pre-training and particle-swarm optimization; 3D CNNs also analyzed bacterial images and time-varying laser speckle.Source: §4.1.1 · pp. 31, 33

4.1.2 Classification based on AlexNet

• AlexNet-based hybrid CNNs classified plankton with 95.83% accuracy, while three-input AlexNet achieved 96.58% in 30 categories and 94.32% in 103.For diatoms, combining original and normalized data produced a best average accuracy of 99.51%.Source: §4.1.2 · p. 33

4.1.3 Classification based on VGGNet

• VGG-16 achieved 80.54% accuracy on SIPPER-77; a transferred parallel VGG-16 achieved 94.98% accuracy and 54.44% F1-score on imbalanced WHOI-Plankton.A comparison of models for bacteria and fungi found Xception and ResNet-50 performed well, and pre-trained weights were not always beneficial.Source: §4.1.3 · p. 34

4.1.4 Classification based on Inception Series

• Inception-V1 achieved about 95% accuracy for five bacteria species, while ImageNet-pretrained Xception reached 97.5% across seven categories.The Xception model was fine-tuned on 740 images.Source: §4.1.4 · pp. 34–35

4.1.5 Classification based on other deep ANNs

• Among other deep networks, DenseNet reached 85% accuracy versus WRN’s 73.2% on seven bacterial categories, and PyramidNet led four plankton CNN baselines.PyramidNet achieved 86.30% accuracy and 41.64% F1-score; fine-tuned ResNet50 achieved 95.44% accuracy and 95.22% F1-score on virus images. MobileNetV2 reached 95.09% average accuracy on 33 bacterial species, while LeNet exceeded 80% on standard-resolution bacteria and yeast images.Source: §4.1.5 · pp. 35–36

4.2.1 Segmentation based on U-Net

• MRFU-Net achieved Dice 87.23%, Jaccard 79.74%, recall 87.65%, accuracy 97.30%, and VOE 20.26% on the 21-category EMDS-5 dataset.The network combines ideas from GoogLeNet and U-Net.Source: §4.2.1 · p. 36
• With center-point annotations dilated into ground truth, a simple U-Net segmented Rift Valley virus in 143 TEM images, achieving Dice 90% and IOU 83.1%.The minimal annotation was proposed to address the expense of image-segmentation annotation.Source: §4.2.1 · p. 36
• For irregular yeast cells, U-Net segmentation combined with watershed using prediction scores addressed adhesion and achieved mean accuracy 94%.U-Net alone could not solve the cell adhesion problem.Source: §4.2.1 · p. 36
• A low-cost U-Net in the CNN-CRF framework achieved pixel-level Dice 87.13%, Jaccard 79.74%, recall 87.12%, accuracy 96.91%, and VOE 20.26%.The framework combines pixel-level segmentation with VGG-16 transfer learning for patch classification and reconstruction; the low-cost U-Net uses less than one-third the memory of U-Net.Source: §4.2.1 · pp. 36–37

4.2.2 Segmentation based on other deep ANNs

• SegNet segmentation of fluorescent yeast-cell microscopy images achieved mIOU 71.72%.The experiment used 6,000 training, 1,200 validation, and 1,200 test samples.Source: §4.2.2 · p. 37
• For tuberculosis bacillus counting from mosaic images, the deepest of three CNNs, CNN1, performed best, achieving accuracy 99.665%.The 5,000 mosaic images each comprised 100 positive and negative patches, approximately half of each.Source: §4.2.2 · p. 37

4.3 Detection Task

• 0.929, 0.973, 0.829, 0.890, and 0.890 were Fast RCNN classification accuracies for five cyanobacteria species; CNN counting achieved R2 0.85 and RMSE 23 cells.Fast RCNN detects and classifies targets before CNN performs counting.Source: §4.3 · p. 38, Figure 17 · p. 39
• 59% accuracy was achieved by Faster RCNN in malaria parasite detection, followed by 98% accuracy from AlexNet classifying targets labeled as non-red blood cells.Both networks were ImageNet-pretrained and fine-tuned on malaria parasite images.Source: §4.3 · p. 38
• 72% was YOLO’s F1-score for detecting diatoms, compared with 10% for RCNN, on a private dataset of nearly 11,000 images in ten categories.Source: §4.3 · p. 39
• 94.8% detection accuracy was reported for a CNN that detects and classifies fungi using a dataset of 40,800 images across five fungus classes and a dirt class [151].The dataset was split into 30,000 training images and 10,800 test images.Source: §4.3 · p. 39

4.4 Feature Extraction Task

• Deep neural networks are used for image feature extraction in addition to classification, segmentation, and detection tasks.The reviewed works commonly pass extracted features to classifiers such as SVM or RF.Source: §4.4 · p. 40
• CNN-derived features support hybrid plankton classification, while DeepLab-VGG-16 features outperform SIFT and simple features for microorganism classification [71].For, SVM using first-hidden-layer features achieved 96.70% accuracy. For [71], pixel feature vectors were combined with global features in a CRF framework.Source: §4.4 · p. 40
• 84% classification accuracy was achieved by DeepSea pre-trained on ISIIS and used as a feature extractor with SVM on the small private dataset.The study used two public datasets and one private dataset, with transfer learning to address the small-dataset problem.Source: §4.4 · pp. 40–41
• Around 96% accuracy was achieved by the Inception-V3-and-SVM bacteria classification system.Pre-trained Inception-V3 was fine-tuned using about 800 training images before serving as the feature extractor for SVM classification.Source: §4.4 · p. 41
• ResNet50 with SVM achieved the best accuracy (94.13%) and recall (94.52%) for plankton identification and enumeration.The framework enhanced ROIs and their local features before extracting CNN features for multi-class SVM classification.Source: §4.4 · p. 41
• 99.5% accuracy was achieved by CNN with Inception-V3 features; the one-/two-round-tuning ensemble scored 0.9527, 0.8826, and 0.9413 accuracy on WHOI, ZooScan, and Kaggle.The ensemble combined models generated by one-round and two-round tuning.Source: §4.4 · p. 41
• 97.989% classification accuracy and 97.986% recall were achieved on in situ plankton by Densenet201+Polar+SVM.On CIFAR-10, the model achieved 94.91% classification accuracy and 94.76% recall.Source: §4.4 · pp. 41–42

4.5 Data Augmentation Task

• Generated images from the proposed framework effectively improve VGG-16 classification performance.The experiment used EMDS-5, with 21 microorganism classes and 20 images plus corresponding ground truth images per class.Source: §4.5 · p. 42
• The GAN framework combines each original image with its ground truth in the same direction to address directional variation and detail loss.These challenges arise because generating images in varied directions is difficult and direct GAN generation on the small dataset may lose details.Source: §4.5 · p. 42
• Color-space transformation increases training data and enables the GAN framework to generate more details.Source: §4.5 · p. 42

4.6 Summary

• Deep-network MIA uses more open-access microorganism datasets, including SIPPER, WHOI-Plankton, and EMDS series.Source: §4.6 · p. 42
• CNN filters automatically extract effective image features, so deep-network MIA usually does not rely on feature engineering and is more flexible than artificial features.Source: §4.6 · p. 42
• Deep networks are usually deeper because convolution layers have fewer parameters than fully connected layers, but they require lots of training data and high-performance computers.Source: §4.6 · p. 42
• Deep-network MIA spans data augmentation, detection, and counting, and more studies focus on optimizing and improving deep networks.Source: §4.6 · p. 42

5.1 Analysis of Methods Based on Classical Neural Networks

• Classical MIA analysis distinguishes classification, segmentation, counting, and feature-extraction tasks, with most methods examined through preprocessing, feature extraction, and network design.Classification commonly uses noise reduction, enhancement, and segmentation before extracting shape, texture, or color features; other tasks may use ANNs for segmentation or feature generation. Classical networks are often shallow because deeper structures were difficult to train with limited hardware.Source: §5.1 · pp. 43, 47
• MLP and RBF are the most widely used classical networks, while feature-engineering dependence makes many methods difficult to transfer across microorganism tasks.Three-layer MLPs are prominent for nonlinear classification and function approximation; a feed-forward MLP applied directly to images in offers an alternative without feature engineering.Source: §5.1 · pp. 47–48, Figure 20 · p. 47

5.2 Analysis of Methods Based on Deep Neural Network

• Deep MIA is centered on CNNs, which learn image features directly and make performance dependent on network structure; VGGNet, Inception, U-Net, AlexNet, and ResNet are widely used public networks.Self-designed CNNs can flexibly combine architectures and training approaches; examples include a network for both fungal detection and classification and LCU-Net for microorganism segmentation.Source: §5.2 · pp. 48–49, Figure 21 · p. 49
• Transfer learning appears in more than 15 MIA papers and can support limited-data tasks or high-level feature extraction.Examples include pretraining on a balanced subset before fine-tuning on the original imbalanced dataset, and comparing plankton versus ImageNet pretraining.Source: §5.2 · pp. 49–50

5.3.1 Potential Direction Based on Existing Methods

• Existing MIA methods could be improved by combining GAN-based augmentation with orientation-normalizing approaches, while importing additional computer-vision techniques.The GAN framework in uses segmentation ground truths to normalize object directions, limiting its use when those labels are unavailable; polar-coordinate images in offer another way to reduce orientation effects.Source: §5.3.1 · p. 50

5.3.2 Potential Direction Based on Dataset Characteristics

• Private or small microorganism datasets motivate few-shot learning and domain adaptation, but current methods still face data-acquisition challenges and a performance gap versus conventional supervised deep learning.The EMDS series has limited samples, and existing work often uses augmentation; few-shot learning and domain adaptation are proposed as ways to address insufficient data.Source: §5.3.2 · p. 51

5.3.3 Potential Direction Based on Advanced Methods

• Transformers are a potential MIA direction because self-attention represents global information, and ViT applies positioned image-patch sequences to classification [39].DETR combines a CNN backbone, encoder-decoder transformer, and feed-forward detection head; its COCO results were comparable to Faster R-CNN. However, pure transformers remain limited by their need for massive training data.Source: §5.3.3 · p. 51

6 Conclusion and Future Work

• The review covers 95 papers on microorganism image analysis, from representative neural networks and classical and deep methods to task-based summaries, method statistics, and future directions.It discusses MLP, VGGNet, Inception, ResNet, U-Net, and YOLO, and analyzes potential development directions including method fusion, dataset characteristics, and methods from other fields.Source: §6 · p. 52
• Most reviewed studies use private datasets, so the review omits a dataset summary and finds no unified standard dataset for evaluating microorganism image analysis methods.Source: §6 · p. 52
• The authors report releasing EMDS-6, EMDS-7, and a SARS-CoV-2 (COVID-19) microscopic image dataset, and call for more open-access datasets and practical ANN-based algorithms.Source: §6 · p. 53